fix(catalog): default openai to gpt-5.5 (matches bundle's extended_thinking=true)

[manojp99]

Why

amplifier-agent ships with a self-contradiction in its own defaults:

Layer	File	Decision
Bundle	src/amplifier_agent_lib/bundle/bundle.md:26	extended_thinking: true on every turn
Catalog	src/amplifier_agent_cli/provider_sources.py:63	openai → default_model: "gpt-4o"
Orchestrator	amplifier-module-loop-streaming:423-424	Forwards extended_thinking as a provider kwarg verbatim
Provider	amplifier-module-provider-openai:1243-1252	Stamps reasoning.effort: "high" on any request carrying extended_thinking=True, with zero model-capability gate

Result with a fresh amplifier-agent run --provider openai:

RuntimeError: Execution failed: InvalidRequestError:
  "Unsupported parameter: 'reasoning.effort' is not supported with this model."

OpenAI 400 on every turn out of the box. The bundle asks for reasoning, the catalog picks a non-reasoning model.

Fix

Switch the catalog's OpenAI default_model from gpt-4o (non-reasoning) to gpt-5.5 (reasoning-capable). amplifier-agent's defaults are now internally consistent — the bundle requests extended_thinking, the catalog ships a model that supports the resulting reasoning.effort parameter.

The fix is one line in src/amplifier_agent_cli/provider_sources.py:63, plus an inline comment explaining why this specific model was chosen.

Why gpt-5.5 specifically

• It accepts the reasoning.effort parameter the bundle's extended_thinking: true ultimately triggers
• It's available across the standard tiers we've verified (/v1/models listing confirms presence on enterprise keys)
• The existing provider has explicit per-effort validation for the gpt-5.5 family (see amplifier-module-provider-openai:118-148 for gpt-5.5-pro validator), so behavior is known and tested
• Cheaper than gpt-5.5-pro for the common case while still reasoning-capable

Consumers who want a different OpenAI model can override via bundle config — this is a default, not a constraint.

What this is NOT

This is not the architecturally correct long-term fix. That belongs in amplifier-module-provider-openai:1243-1252 — gate the extended_thinking → reasoning.effort injection on _model_may_reason(model_name), mirroring the discipline already present at lines 1239-1242. That fix closes the bug for every consumer (amplifier-agent, amplifier-app-cli, openclaw, anything else using the provider) regardless of which model they default to. It will be filed against that repo as a separate PR.

This PR is the immediate-unblock so amplifier-agent users get a working OpenAI default now, without waiting for the provider fix to merge + propagate.

What needs to rebuild downstream

Component	Action
amplifier-agent-ts (wrapper)	none — catalog content isn't a wire contract
Protocol version	unchanged (0.2.0)
pyproject.toml engine version	not bumped — forward-fix on engine-v0.3.0; the tag will be moved to this merge commit
Bundle (bundle.md)	unchanged
bundle.version	unchanged (1.2.1) — no bundle behavior change, so no cache invalidation needed
nanoclaw Dockerfile AMPLIFIER_AGENT_REF	unchanged (engine-v0.3.0); image rebuild picks up the moved tag's new SHA
Other consumers (amplifier-app-cli, openclaw)	unaffected — their own catalogs ship the same defect; they get the principled fix when the provider gate lands

Verification

Build + manual amplifier-agent run --provider openai -y "hi" in the rebuilt nanoclaw container will exercise the new default. Expected: bundle sends extended_thinking=true, provider injects reasoning.effort: "high", OpenAI accepts the request against gpt-5.5, model returns a reply.